An agent that remembers everything drowns in its own history. Forgetting on purpose is the discipline that keeps memory useful — and it's the one most teams skip.
Memory eviction is the unpopular half of memory management: deciding what to throw away. Teams build the store and never the eviction policy, and the store grows until retrieval degrades and cost balloons. Forgetting is the feature that keeps remembering affordable.
Sharpen the Saw: Eviction Is Maintenance
Eviction is saw-sharpening — the periodic, unglamorous practice that keeps the system working. A memory store without eviction is a store that gets slower, noisier, and more expensive over time. Eviction is what keeps it fast, precise, and cheap.
- Run eviction on a cadence, not on demand.
- Evict the stale, the one-off, the contradicted.
- Archive the rare-but-valuable; don't keep it in the hot path.
The Cashflow Lens
Memory has a cost: storage, retrieval latency, and the quality cost of noise. An unbounded store costs more over time, not less. Eviction is the discipline that keeps the cost bounded — the store stays small, retrieval stays fast, and the agent stays precise.
What to Evict
- One-off events that never recurred.
- Stale facts that have been superseded.
- Contradicted entries that were resolved.
- High-noise, low-signal history.
What to Keep
- Recurring patterns (the lessons).
- Stable facts and preferences.
- Recent episodic context (within a window).
- Anything audit-relevant, archived and out of the hot path.
Conclusion
Forgetting on purpose is the discipline that keeps agent memory useful. Evict on a cadence — the stale, the one-off, the contradicted — and the store stays small, fast, and precise. The agent that remembers everything drowns; the agent that forgets well, remembers well.
About FACTA
FACTA helps startups and growth-stage teams turn AI into production systems that keep running — not demos that impress once.
We design the architecture around the parts that actually break under real usage: tooling you own, credentials you control, failover, cost controls, observability. The boring infrastructure that keeps a system alive after launch.
Led by Matías Baglieri and Carolina Fogliato, we focus on one thing:
AI leadership that builds. Not just advises.
Ask us what your agent's memory policy is for throwing things away.
We'll tell you what eviction would save. See memory consolidation for what to keep.
Explore AI Automation
